Neural Net from Scratch
While taking Northwestern University’s CS469 Machine Learning and Artificial Intelligence for Robotics class, I was tasked with selecting a learning aim applicable to a given dataset and a machine learning model suitable to learn the aim.
Neural networks and various machine learning models are becoming increasingly popular to solve complex learning aims. In this assignment, I developed and optimized a neural network to learn the mobile robots motion model. The inputs to the model were the robots current state at t (x,y,heading), linear and angular velocities (v,w), and the change in time (dt), and the model output the new state at t+dt x,y,heading. The data used was from this experiment and is included in the provided Github Repository below.
The model’s weights and biases were learned using gradient descent. I also applied mini batching to decrease the model’s training time. More of these techniques are broken down in the assignment writeup that can be found on Github.